Dakera vs Weaviate: modules you assemble vs an engine that ships integrated
Weaviate is an open-source vector database that connects to embedding, generative and reranker model providers through integrations, and it now offers Engram, a managed memory service. Dakera is one self-hosted binary whose memory pipeline is already wired together. The question is how much of the pipeline you want to choose, and how much you want shipped.
Reviewed October 2026
Assembled, or integrated
With Weaviate you choose a provider for each stage of retrieval, and your application owns the memory logic around it. That flexibility is real: the choice runs from hosted APIs to locally hosted models. Dakera makes those choices once, on-device, and puts the memory logic in the server.
Where Weaviate leads
Weaviate is a broad platform, and on the database side it is generous with options.
- Integrations: embedding, generative and reranker integrations across providers such as OpenAI, Cohere, Google and Anthropic, plus locally hosted Hugging Face Transformers and Ollama.
- Index and compression choices: HNSW, flat, dynamic and the disk-based HFresh index; product, binary, scalar and rotational quantization.
- Multi-vector: ColBERT-style multi-vector embeddings since Weaviate 1.29. Dakera offers multi-vector records and late interaction as opt-in v0.12.0 features, so this is parity rather than a gap.
- Hybrid search: relative-score and ranked fusion with an alpha weight.
- Agents: an MCP server, generally available since 1.38, and a Query Agent.
The memory layer: Engram and Dakera
Weaviate describes Engram as a fully managed memory and context service that extracts, consolidates and shares memories across agents, scoped by project and user. Dakera covers the same ground inside your own deployment.
| Capability | Weaviate database | Engram | Dakera |
|---|---|---|---|
| Runs on your infrastructure | Yes | Managed service | Yes |
| Extracts memories from conversationsDakera extracts entities on-device; no LLM call | No | Yes | Partly |
| Consolidates related memoriesDakera: deduplication and consolidation | No | Yes | Yes |
| Memory shared across agentsDakera: edges are private to their agent by default | No | Yes | Yes |
| Scoped by user, project or agentDakera: agents, sessions, namespaces and scoped API keys | Partly | Yes | Yes |
| Decay curves and spaced-repetition TTL | No | Not stated | Yes |
| Knowledge graph view and recall inspector | No | Not stated | Yes |
"Not stated": the Engram documentation does not say. "No" in the database column means the database layer does not offer it; you build it in your application.
Retrieval you can inspect
Dakera's recall is one pipeline: embed on-device, retrieve with BM25 and HNSW, fuse with RRF, then rescore with a cross-encoder reranker. v0.12.0 adds rerank_candidates to bound the reranker per request, and every response carries a rerank_report. The Dashboard shows the same data per memory.
Languages and media
| Weaviate | Dakera v0.12.0 | |
|---|---|---|
| Multilingual | Choose a multilingual embedding integration | bge-m3 in one space for 100+ languages; stemming in 18 languages plus CJK bigrams; per-request lang. Details |
| Multimodal | Image and other media through multimodal integrations | Attachments, on-device speech to text (5 Whisper models) and visual recall over document pages. Details |
| Air-gapped | Possible with locally hosted models | Images ship their default models; dakera models list|pull|prune manages the rest |
When to choose Weaviate, and when Dakera
Choose Weaviate if
- You want to pick the embedding, reranker and generative provider per workload, or swap them over time.
- You need generative search and the Query Agent next to your retrieval.
- You want a choice of disk-based or compressed indexes for large collections.
- You prefer one vendor for a managed database and a managed memory service (Engram).
- A BSD-3-Clause core matters; enterprise features sit under a separate commercial licence.
Choose Dakera if
- You want the pipeline shipped: embeddings, reranking, entities, decay and sessions in one binary.
- Memory must stay on your infrastructure, with no external model API in the memory path.
- You need multilingual or multimodal memory without choosing and wiring providers.
- You want a bill that does not depend on a provider, a token count or a cloud plan.
Weaviate Cloud has a free tier (100,000 objects) and a Flex plan from $45 a month. Prefer managed? Dakera Cloud is coming. Join the waitlist.
Frequently asked questions
Does Weaviate have agent memory?
Yes. Weaviate describes Engram as a fully managed memory and context service built on its vector database. The Weaviate database itself is a general vector store, where memory logic lives in your application.
Can Weaviate run without external model APIs?
Yes. It supports locally hosted models through Hugging Face Transformers and Ollama, which lets it run air-gapped. Dakera takes the same position by default: its images ship the models they need.
How do I move memories from Weaviate to Dakera?
Export your objects with their text, start Dakera with the Quickstart command and store them with store_memory or the batch endpoint. Dakera embeds on-device, so vectors are not carried across.
Is the LoCoMo figure comparable?
Dakera reports Recall@20, whether the answer is among the top 20 retrieved memories, with no LLM in the retrieval path. It is a memory-retrieval measure and is not set against Weaviate's database benchmarks.
Run Dakera on your own infrastructure.
One binary with its models baked in and no per-call fees. Start it with one Docker command and see it for yourself.
Prefer managed hosting? Join the Dakera Cloud waitlist →